Automated musical instrument recognition

Music information retrieval, or MIR, is the science of extracting musical features from audio signals. One such feature extraction is the ability to determine which instrument is playing, given a monophonic musical signal. The problem has been dealt with successfully mostly with different machine learning algorithms, trained with multiple feature representations of the audio. In recent years, there has been a growing interest for deep learning within several problem domains, also within MIR. The idea of letting the model learn the distinguishing features instead of extracting them by hand is attractive in a field where most representations are arduous for manual interpretation. This thesis investigates the task of automatic instrument recognition by training a convolutional neural network with a single spectral representation of sound. Two different models are tested, each fed with three different spectral representations. The results show average accuracies of around 65% for 16 instruments when tested with a unseen data containing a mix of real life recordings and isolated notes.